Pan Mu, Zhu Liu, Yaohua Liu, Risheng Liu, and Xin Fan
Video deraining is an important issue for outdoor vision systems and has been investigated extensively. However, designing optimal architectures by the aggregating model formation and data distribution is a challenging task for video deraining. In this paper, we develop a model guided triple-level optimization framework to deduce network architecture with cooperating optimization and auto-searching mechanism, named Triple-level Model Inferred Cooperating Searching (TMICS), for dealing with various video rain circumstances. In particular, to better explore inter-frame information from videos, we first introduce a macroscopic structure searching scheme that searches from Optical Flow Module (OFM) and Temporal Grouping Module (TGM) to help restore the latent frame. In addition, existing methods cannot preserve details and structure when removing rain streaks. To overcome the problems, we then design a collaborative structure for video deraining based on the proposed optimization model. This structure includes Dominate Network Architecture (DNA) and Companionate Network Architecture (CNA) and is cooperated by introducing an Attention-based Averaging Scheme (AAS). To obtain suitable task-specific architectures (i.e., DNA and CNA), we apply the differentiable neural architecture search from a compact candidate set of task-specific operations to discover desirable rain streaks removal architectures automatically. Extensive experiments on various datasets demonstrate that our model shows significant improvements in fidelity and temporal consistency over the state-of-the-art works.
- Linux or Windows
- Python 3
- NVIDIA GPU + CUDA cuDNN
- PyTorch 1.2
- Video supplementary material: Baiduyun(code: o09j)
- Models:Baiduyun(code: 72vy)